The Decline of Social Media Censorship and the Rise of Self-Censorship after the 2016 Failed Turkish Coup

In this work, we examine the effect of the 2016 Turkish coup on social media censorship, both by the government ordering Twitter to conduct censorship and as well by people removing their own tweets. We compared 5.5M tweets collected from Turkey pre-coup to 8.5M tweets collected post-coup. Although self-censorship of the press is not a novel practice following past military coups in Turkey, in this work we examine and quantify social media self-censorship, and empirically compare its effect relative to government-implemented censorship of social media. Our measurements following the coup show a 72% decline in publicly identifiable government-censored tweets. We attribute this, in part, to an estimated 43% decline in overall Twitter usage in Turkey and in part to users’ self-censorship. Supporting this theory, we detected that 41% of all users in our pre-coup dataset voluntarily removed 18% of their old tweets by either switching their accounts to protected mode, deleting their accounts, or deleting some tweets. Using NLP and graph metrics, we identify a new focus of Turkish government censorship on the Gülen movement. Our analysis show pro-Gülen tweets being widely self-censored. Additionally, we detected 40% more publicly-accessible anti-Gülen tweets. Unlike activists who regularly tweet political content, and are more likely to be censored by the government, we found that self-censoring users appear to be more typical users who normally post neutral tweets, and only 6% political tweets on average.